AI
Autoresearch: The feedback loop behind self
Introspection, a new company founded by ex-xAI engineers Roland Gavrilescu and Julian Bright, is building infrastructure for self-improving AI agent systems
Key takeaways
- Autoresearch involves building an "outer loop" where agents help maintain and improve the primary system using feedback signals, evals, and human input.
- The loop itself is the product, shifting focus from models to harnesses to loops.
- Agent recipes are a portable format that brings together evals, judges, signal processing, and human expertise to enable agents to iterate like a research laboratory.
- Introspection’s Pi framework separates the agent loop from its extensions and configuration, making agents portable and easy to deploy.
- Humans remain central to the system, acting as a source of signals and tacit knowledge during early loops.
Introspection, a new company founded by ex-xAI engineers Roland Gavrilescu and Julian Bright, is building infrastructure for self-improving AI agent systems called "autoresearch." The company introduces the concept of agent "recipes" - portable, provider-agnostic containers that encode human expertise, evals, judges, and feedback signals to enable agents to iteratively improve systems over time. Introspection's framework, Pi, is compared to Linux for agent harnesses, and the company focuses on vertical SaaS and services businesses, aiming to bring frontier AI lab-level agent technology to production environments while maintaining data ownership and avoiding vendor lock-in. The system is designed with humans in the loop, using Git as an audit log and initially targeting software engineers in vertical SaaS companies.
In their words
“The loop is the product. We have moved from focusing on models, to harnesses, and now to loops.”
“You begin with a baseline and then record how each signal produced a new judge, embedded new human expertise or led you to introduce a different model.”
By the numbers
- 2026
- year Introspection was founded
- 3
- patterns forming autoresearch blueprint
How it unfolded
- Gavrilescu and Bright worked at xAI
- Introspection founded
- AI Engineer World’s Fair held
Turn stories like this into views
Ravenclip finds the AI news, makes the video, and posts it before attention moves on.
Common questions
- What happened with Introspection?
- Autoresearch involves building an "outer loop" where agents help maintain and improve the primary system using feedback signals, evals, and human input.
- Where can I read the original report?
- Read the full report at latent_space.
